Supercomputers excel at highly calculation-intensive tasks, such asmolecular modeling and large-scale simulations, and have enabledsignificant scientific breakthroughs.
Yet supercomputers themselves are subject to technologicaladvancements and redesigns that allow them to keep pace with thescience they support.
The current vision of future supercomputers calls for them tobe very heterogeneous--for example, rather than a central processingunit (CPU) with memory, disk and interconnect, the CPU will containcores of smaller CPUs making up a larger whole--and have differenttypes of processors, such as vectors and field programmable gate arrays(FPGAs). The location and type of memory will be more complex as well.
High performance components--encapsulated chunks of softwarethat perform specific tasks--will be coupled to a dynamic frameworkthat allows the scientists and the software to dynamically determinethe algorithms or modifications to algorithms that will perform well ona particular architecture.
Multiple levels of parallelism will be explored, includingparallelism at the component level, parallelism within the component,parallelism within a subroutine and threading.
These supercomputers of the future will provide orders ofmagnitude more computing power, but their increasing complexity alsorequires experts in computational science, mathematics and computerscience working together to develop the software needed for thescience.
Pacific Northwest National Laboratory researcher Theresa Windus will be presenting her results at 9:45 a.m., Tuesday, Aug. 30.
Cite This Page: